“Xiaomi achieved high inference speed for its MiMo-V2.5-Pro-UltraSpeed model through techniques including FP4 quantization, DFlash speculative decoding, and using TileRT software from Tile AI.”